FAIRM: Learning invariant representations for algorithmic fairness and domain generalization with minimax optimality.